Speakers
Description
Accelerators in GNU Radio have been previously limited to requiring the selection of an accelerator tor framework/block at design time. In this talk and associated paper, we present a preliminary investigation into supporting two new capabilities with GNU Radio: first, we illustrate the ability to execute GNU Radio blocks across a variety of heterogeneous accelerators (FPGA and GPU). Second, we demonstrate that we are able to dynamically schedule these blocks across our pool of accelerators using easily-customizable scheduling policies. We do this via a combination of out-of-tree modules and by embedding GNU Radio itself as an application in a heterogeneous runtime called CEDR (Mack et al., 2022). The CEDR ecosystem provides a productive environment for researching the combined challenges of application design, systems software, and hardware prototyping for heterogeneous systems, and namely, it provides a flexible intelligent scheduling (IS) interface by which the user can easily adjust or prioritize how tasks are dispatched at runtime to the various accelerators present on a given system. The IS integrates a variety of runtime schedulers that optimize for metrics such as application performance and minimum scheduling algorithm overheads. The IS in tandem with CEDR improves the performance of applications through dynamic scheduling by efficiently utilizing the resources at runtime.
We demonstrate GNU Radio running on heterogeneous hardware using CEDR across both Xilinx Zynq Ultrascale+ ZCU102 and Nvidia Jetson AGX Xavier systems. We choose three applications that can leverage FFT acceleration: we implement Pulse Doppler radar in GNU Radio, deploy in CEDR and execute it along with non-GNU Radio-based WiFi-TX and Synthetic Aperture Radar applications. We find that, when GNU Radio shares the system with other unrelated applications, our integrated IS dispatches FFT tasks to the accelerator on both the FPGA and GPU platforms along with the CPU cores fairly without compromising target throughput for each application. In summary, we show that running the GNU Radio runtime and applications inside/with CEDR enables better scheduling options and easier accelerator access by eliminating the need for users to partition their workloads at design time. We believe this is a stepping stone in broadening GNU Radio’s support for heterogeneous execution and enabling it to hook more flexibly into a variety of scheduling heuristics.
https://github.com/UA-RCL/CEDR
https://ua-rcl.github.io/CEDR/
| Talk Length | 30 Minutes |
|---|---|
| Acknowledge | Acknowledge In-Person |
